Linklaters

Law firms deploying AI document review tools consistently encounter a ceiling on accuracy: most systems can extract and analyze only a portion of contract content before the unpredictability of unstructured legal language causes errors or gaps. Linklaters identified that this limitation is not primarily a failure of the AI models themselves but a data quality problem originating at the moment contracts are drafted. Legal documents are assembled from heterogeneous sources — bespoke clauses, negotiated language, inconsistent formatting — making automated extraction inherently unreliable. For a global firm handling complex derivatives and cross-border transactions at scale, accepting 70% automated review accuracy was operationally insufficient and left meaningful attorney time trapped in manual contract analysis.

Linklaters built Nakhoda, an in-house legal AI platform developed by a dedicated team of approximately 25 people since 2015, to address the problem at its source rather than downstream. Rather than applying natural language processing to legacy unstructured documents, Nakhoda restructures the contracting process itself: documents are created within a controlled digital environment using shared clause banks, agreed templates, and a standardized contracting workflow, making them machine-readable by design. The first major deployment targeted ISDA derivatives contracts — a rules-heavy, well-defined contract type suited to templated methodology. Linklaters partnered directly with the global ISDA rules body to build the platform around an incoming wave of margin-related derivatives regulations, embedding counterparty-finding capabilities alongside automated review. The architecture ensures that NLP extraction operates on predictable, structured inputs rather than open-ended free text.

The ISDA platform demonstrates a measurable improvement in automated review completeness when contracts are originated within Nakhoda's structured environment. Because the extraction layer already has a high-confidence model of what each document contains — clause type, position, and permitted variation — NLP-driven review approaches near-complete coverage rather than the partial results typical of legacy document analysis. Qualitative outcomes include:

• Counterparty discovery and matching integrated directly into the contracting workflow
• A contracting environment co-developed with the ISDA rules body, positioning the platform ahead of incoming margin regulation changes
• Validation that fixing legal data at the point of creation is more effective than improving extraction models alone

The platform is being extended to additional contract types as the methodology matures.

• Fix the data at creation, not extraction: AI review accuracy is constrained by source document quality — investing in structured drafting environments yields more reliable results than tuning models against unstructured legacy contracts.
• Start with rules-dense contract types: ISDA derivatives provided a well-bounded, regulation-driven use case that let Nakhoda prove the model before tackling more variable contract categories.
• Shared infrastructure multiplies value: Clause banks and agreed templates that all counterparties work from reduce negotiation variability and make automated review substantially more accurate.
• In-house development requires long-term commitment: Building a 25-person legal AI team over four years reflects an organizational investment level most firms should factor into build-vs-buy decisions.